Customer Clustering On Mobile Internet Online Behavior

The prosperity of mobile Internet and4G technology have given birth to the mobile online services and products. Customers nowadays are accustomed to use mobile terminal for online activities, such as shopping and instant messaging, which leaves millions of personal data. All these data, including the service type, service usage time, service usage frequency etc., describe customers’ service preferences as well as the behavior patterns comprehensively. Besides, along with the development of big data technology, companies begin to learn the way customer’s data impacting their service and product strategies. With the analysis of customer behavior, we can learn more about different kinds of customers and formulate marketing or service planning according to their characteristics.Most of previous studies investigate the customer information in single dimension, but customer’s information is multidimensional. For instance, even though different customers have similar frequency of applying some mobile applications, some of them would like to use it on daytime while others prefer to apply it in the midnight. Therefore, only investigate customers in frequency domain would leave out some critical information. Our study propose a customer clustering method, which segment customer behavior information into periodical time sequences and then fit the probability density model in both temporal and frequency domains. The probability density models in different dimensions would describe the customer usage preferences from different perspectives. After which, we compute the dissimilarity distance of each customer and build a similarity matrix for the follow-up clustering models. When building the clustering model, we firstly apply the classic algorithms such as the hierarchy algorithm, k-medoid algorithm and spectral clustering algorithm. In addition, we also take advantage of the community detection models. After applying all these methods on the customer data, we build a clustering ensemble model to improve the clustering accuracy and robustness. The clustering results would inspire companies for both the product design and strategic planning.